2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Simple copy minimization heuristics.
23 * @author Christian Wuerdig
26 * This is the C implementation of the mst algorithm
27 * originally written in Java by Sebastian Hack.
28 * (also known as "heur3" :)
29 * Performs simple copy minimization.
33 #define DISABLE_STATEV
40 #include "raw_bitset.h"
41 #include "irnodemap.h"
55 #include "becopyopt_t.h"
59 #define COL_COST_INFEASIBLE DBL_MAX
60 #define AFF_NEIGHBOUR_FIX_BENEFIT 128.0
61 #define NEIGHBOUR_CONSTR_COSTS 64.0
66 #define DBG_AFF_CHUNK(env, level, chunk) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_aff_chunk((env), (chunk)); } while (0)
67 #define DBG_COL_COST(env, level, cost) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_col_cost((env), (cost)); } while (0)
69 static firm_dbg_module_t *dbg = NULL;
73 #define DBG_AFF_CHUNK(env, level, chunk)
74 #define DBG_COL_COST(env, level, cost)
79 #define REAL(C) (C ## f)
81 static unsigned last_chunk_id = 0;
82 static int recolor_limit = 7;
83 static double dislike_influence = REAL(0.1);
85 typedef struct col_cost_t {
93 typedef struct aff_chunk_t {
94 const ir_node **n; /**< An ARR_F containing all nodes of the chunk. */
95 const ir_node **interfere; /**< An ARR_F containing all inference. */
96 int weight; /**< Weight of this chunk */
97 unsigned weight_consistent : 1; /**< Set if the weight is consistent. */
98 unsigned deleted : 1; /**< For debugging: Set if the was deleted. */
99 unsigned id; /**< An id of this chunk. */
102 col_cost_t color_affinity[1];
108 typedef struct aff_edge_t {
109 const ir_node *src; /**< Source node. */
110 const ir_node *tgt; /**< Target node. */
111 int weight; /**< The weight of this edge. */
114 /* main coalescing environment */
115 typedef struct co_mst_env_t {
116 int n_regs; /**< number of regs in class */
117 int k; /**< number of non-ignore registers in class */
118 bitset_t *allocatable_regs; /**< set containing all global ignore registers */
119 ir_nodemap map; /**< phase object holding data for nodes */
121 pqueue_t *chunks; /**< priority queue for chunks */
122 list_head chunklist; /**< list holding all chunks */
123 be_ifg_t *ifg; /**< the interference graph */
124 copy_opt_t *co; /**< the copy opt object */
125 unsigned chunk_visited;
126 col_cost_t **single_cols;
129 /* stores coalescing related information for a node */
130 typedef struct co_mst_irn_t {
131 const ir_node *irn; /**< the irn this information belongs to */
132 aff_chunk_t *chunk; /**< the chunk this irn belongs to */
133 bitset_t *adm_colors; /**< set of admissible colors for this irn */
134 ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */
135 int n_neighs; /**< length of the interfering neighbours array. */
136 int int_aff_neigh; /**< number of interfering affinity neighbours */
137 int col; /**< color currently assigned */
138 int init_col; /**< the initial color */
139 int tmp_col; /**< a temporary assigned color */
140 unsigned fixed : 1; /**< the color is fixed */
141 struct list_head list; /**< Queue for coloring undo. */
142 real_t constr_factor;
146 * In case there is no phase information for irn, initialize it.
148 static co_mst_irn_t *co_mst_irn_init(co_mst_env_t *env, const ir_node *irn)
150 co_mst_irn_t *res = OALLOC(&env->obst, co_mst_irn_t);
152 const arch_register_req_t *req;
153 neighbours_iter_t nodes_it;
161 res->int_neighs = NULL;
162 res->int_aff_neigh = 0;
163 res->col = arch_register_get_index(arch_get_irn_register(irn));
164 res->init_col = res->col;
165 INIT_LIST_HEAD(&res->list);
167 DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn));
169 /* set admissible registers */
170 res->adm_colors = bitset_obstack_alloc(&env->obst, env->n_regs);
172 /* Exclude colors not assignable to the irn */
173 req = arch_get_irn_register_req(irn);
174 if (arch_register_req_is(req, limited)) {
175 rbitset_copy_to_bitset(req->limited, res->adm_colors);
177 bitset_set_all(res->adm_colors);
180 /* exclude global ignore registers as well */
181 bitset_and(res->adm_colors, env->allocatable_regs);
183 /* compute the constraint factor */
184 res->constr_factor = (real_t) (1 + env->n_regs - bitset_popcount(res->adm_colors)) / env->n_regs;
186 /* set the number of interfering affinity neighbours to -1, they are calculated later */
187 res->int_aff_neigh = -1;
189 /* build list of interfering neighbours */
191 be_ifg_foreach_neighbour(env->ifg, &nodes_it, irn, neigh) {
192 if (!arch_irn_is_ignore(neigh)) {
193 obstack_ptr_grow(&env->obst, neigh);
197 res->int_neighs = (ir_node**)obstack_finish(&env->obst);
202 static co_mst_irn_t *get_co_mst_irn(co_mst_env_t *env, const ir_node *node)
204 co_mst_irn_t *res = ir_nodemap_get(co_mst_irn_t, &env->map, node);
206 res = co_mst_irn_init(env, node);
207 ir_nodemap_insert(&env->map, node, res);
212 typedef int decide_func_t(const co_mst_irn_t *node, int col);
217 * Write a chunk to stderr for debugging.
219 static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c)
223 if (c->weight_consistent)
224 ir_fprintf(stderr, " $%d ", c->weight);
225 ir_fprintf(stderr, "{");
226 for (i = 0, l = ARR_LEN(c->n); i < l; ++i) {
227 const ir_node *n = c->n[i];
228 ir_fprintf(stderr, " %+F,", n);
230 ir_fprintf(stderr, "}");
234 * Dump all admissible colors to stderr.
236 static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node)
240 if (bitset_popcount(node->adm_colors) < 1)
241 fprintf(stderr, "no admissible colors?!?");
243 bitset_foreach(node->adm_colors, idx) {
244 ir_fprintf(stderr, " %zu", idx);
250 * Dump color-cost pairs to stderr.
252 static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost)
255 for (i = 0; i < env->n_regs; ++i)
256 fprintf(stderr, " (%d, %.4f)", cost[i].col, cost[i].cost);
259 #endif /* DEBUG_libfirm */
261 static inline int get_mst_irn_col(const co_mst_irn_t *node)
263 return node->tmp_col >= 0 ? node->tmp_col : node->col;
267 * @return 1 if node @p node has color @p col, 0 otherwise.
269 static int decider_has_color(const co_mst_irn_t *node, int col)
271 return get_mst_irn_col(node) == col;
275 * @return 1 if node @p node has not color @p col, 0 otherwise.
277 static int decider_hasnot_color(const co_mst_irn_t *node, int col)
279 return get_mst_irn_col(node) != col;
283 * Always returns true.
285 static int decider_always_yes(const co_mst_irn_t *node, int col)
292 /** compares two affinity edges by its weight */
293 static int cmp_aff_edge(const void *a, const void *b)
295 const aff_edge_t *e1 = (const aff_edge_t*)a;
296 const aff_edge_t *e2 = (const aff_edge_t*)b;
298 if (e2->weight == e1->weight) {
299 if (e2->src->node_idx == e1->src->node_idx)
300 return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx);
302 return QSORT_CMP(e2->src->node_idx, e1->src->node_idx);
304 /* sort in descending order */
305 return QSORT_CMP(e2->weight, e1->weight);
308 /** compares to color-cost pairs */
309 static __attribute__((unused)) int cmp_col_cost_lt(const void *a, const void *b)
311 const col_cost_t *c1 = (const col_cost_t*)a;
312 const col_cost_t *c2 = (const col_cost_t*)b;
313 real_t diff = c1->cost - c2->cost;
320 return QSORT_CMP(c1->col, c2->col);
323 static int cmp_col_cost_gt(const void *a, const void *b)
325 const col_cost_t *c1 = (const col_cost_t*)a;
326 const col_cost_t *c2 = (const col_cost_t*)b;
327 real_t diff = c2->cost - c1->cost;
334 return QSORT_CMP(c1->col, c2->col);
338 * Creates a new affinity chunk
340 static inline aff_chunk_t *new_aff_chunk(co_mst_env_t *env)
342 aff_chunk_t *c = XMALLOCF(aff_chunk_t, color_affinity, env->n_regs);
343 c->n = NEW_ARR_F(const ir_node *, 0);
344 c->interfere = NEW_ARR_F(const ir_node *, 0);
346 c->weight_consistent = 0;
348 c->id = ++last_chunk_id;
350 list_add(&c->list, &env->chunklist);
355 * Frees all memory allocated by an affinity chunk.
357 static inline void delete_aff_chunk(aff_chunk_t *c)
360 DEL_ARR_F(c->interfere);
367 * binary search of sorted nodes.
369 * @return the position where n is found in the array arr or ~pos
370 * if the nodes is not here.
372 static inline int nodes_bsearch(const ir_node **arr, const ir_node *n)
374 int hi = ARR_LEN(arr);
378 int md = lo + ((hi - lo) >> 1);
391 /** Check if a node n can be found inside arr. */
392 static int node_contains(const ir_node **arr, const ir_node *n)
394 int i = nodes_bsearch(arr, n);
399 * Insert a node into the sorted nodes list.
401 * @return 1 if the node was inserted, 0 else
403 static int nodes_insert(const ir_node ***arr, const ir_node *irn)
405 int idx = nodes_bsearch(*arr, irn);
408 int i, n = ARR_LEN(*arr);
411 ARR_APP1(const ir_node *, *arr, irn);
416 for (i = n - 1; i >= idx; --i)
425 * Adds a node to an affinity chunk
427 static inline void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node)
431 if (! nodes_insert(&c->n, node->irn))
434 c->weight_consistent = 0;
437 for (i = node->n_neighs - 1; i >= 0; --i) {
438 ir_node *neigh = node->int_neighs[i];
439 nodes_insert(&c->interfere, neigh);
444 * Check if affinity chunk @p chunk interferes with node @p irn.
446 static inline int aff_chunk_interferes(const aff_chunk_t *chunk, const ir_node *irn)
448 return node_contains(chunk->interfere, irn);
452 * Check if there are interference edges from c1 to c2.
454 * @param c2 Another chunk
455 * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise.
457 static inline int aff_chunks_interfere(const aff_chunk_t *c1, const aff_chunk_t *c2)
464 /* check if there is a node in c2 having an interfering neighbor in c1 */
465 for (i = ARR_LEN(c2->n) - 1; i >= 0; --i) {
466 const ir_node *irn = c2->n[i];
468 if (node_contains(c1->interfere, irn))
475 * Returns the affinity chunk of @p irn or creates a new
476 * one with @p irn as element if there is none assigned.
478 static inline aff_chunk_t *get_aff_chunk(co_mst_env_t *env, const ir_node *irn)
480 co_mst_irn_t *node = get_co_mst_irn(env, irn);
485 * Let chunk(src) absorb the nodes of chunk(tgt) (only possible when there
486 * are no interference edges from chunk(src) to chunk(tgt)).
487 * @return 1 if successful, 0 if not possible
489 static int aff_chunk_absorb(co_mst_env_t *env, const ir_node *src, const ir_node *tgt)
491 aff_chunk_t *c1 = get_aff_chunk(env, src);
492 aff_chunk_t *c2 = get_aff_chunk(env, tgt);
495 DB((dbg, LEVEL_4, "Attempt to let c1 (id %u): ", c1 ? c1->id : 0));
497 DBG_AFF_CHUNK(env, LEVEL_4, c1);
499 DB((dbg, LEVEL_4, "{%+F}", src));
501 DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %u): ", c2 ? c2->id : 0));
503 DBG_AFF_CHUNK(env, LEVEL_4, c2);
505 DB((dbg, LEVEL_4, "{%+F}", tgt));
507 DB((dbg, LEVEL_4, "\n"));
512 /* no chunk exists */
513 co_mst_irn_t *mirn = get_co_mst_irn(env, src);
516 for (i = mirn->n_neighs - 1; i >= 0; --i) {
517 if (mirn->int_neighs[i] == tgt)
521 /* create one containing both nodes */
522 c1 = new_aff_chunk(env);
523 aff_chunk_add_node(c1, get_co_mst_irn(env, src));
524 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
528 /* c2 already exists */
529 if (! aff_chunk_interferes(c2, src)) {
530 aff_chunk_add_node(c2, get_co_mst_irn(env, src));
534 } else if (c2 == NULL) {
535 /* c1 already exists */
536 if (! aff_chunk_interferes(c1, tgt)) {
537 aff_chunk_add_node(c1, get_co_mst_irn(env, tgt));
540 } else if (c1 != c2 && ! aff_chunks_interfere(c1, c2)) {
543 for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx)
544 aff_chunk_add_node(c1, get_co_mst_irn(env, c2->n[idx]));
546 for (idx = 0, len = ARR_LEN(c2->interfere); idx < len; ++idx) {
547 const ir_node *irn = c2->interfere[idx];
548 nodes_insert(&c1->interfere, irn);
551 c1->weight_consistent = 0;
553 delete_aff_chunk(c2);
556 DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n"));
560 DB((dbg, LEVEL_4, " ... absorbed\n"));
565 * Assures that the weight of the given chunk is consistent.
567 static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c)
569 if (! c->weight_consistent) {
573 for (i = 0; i < env->n_regs; ++i) {
574 c->color_affinity[i].col = i;
575 c->color_affinity[i].cost = REAL(0.0);
578 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
579 const ir_node *n = c->n[idx];
580 const affinity_node_t *an = get_affinity_info(env->co, n);
581 co_mst_irn_t *node = get_co_mst_irn(env, n);
584 if (node->constr_factor > REAL(0.0)) {
585 bitset_foreach (node->adm_colors, col)
586 c->color_affinity[col].cost += node->constr_factor;
590 co_gs_foreach_neighb(an, neigh) {
591 const ir_node *m = neigh->irn;
593 if (arch_irn_is_ignore(m))
596 w += node_contains(c->n, m) ? neigh->costs : 0;
601 for (i = 0; i < env->n_regs; ++i)
602 c->color_affinity[i].cost *= (REAL(1.0) / ARR_LEN(c->n));
605 // c->weight = bitset_popcount(c->nodes);
606 c->weight_consistent = 1;
611 * Count the number of interfering affinity neighbours
613 static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an)
615 const ir_node *irn = an->irn;
616 const co_mst_irn_t *node = get_co_mst_irn(env, irn);
619 co_gs_foreach_neighb(an, neigh) {
620 const ir_node *n = neigh->irn;
623 if (arch_irn_is_ignore(n))
626 /* check if the affinity neighbour interfere */
627 for (i = 0; i < node->n_neighs; ++i) {
628 if (node->int_neighs[i] == n) {
639 * Build chunks of nodes connected by affinity edges.
640 * We start at the heaviest affinity edge.
641 * The chunks of the two edge-defining nodes will be
642 * merged if there are no interference edges from one
643 * chunk to the other.
645 static void build_affinity_chunks(co_mst_env_t *env)
647 nodes_iter_t nodes_it;
648 aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0);
651 aff_chunk_t *curr_chunk;
654 /* at first we create the affinity edge objects */
655 be_ifg_foreach_node(env->ifg, &nodes_it, n) {
656 int n_idx = get_irn_idx(n);
660 if (arch_irn_is_ignore(n))
663 n1 = get_co_mst_irn(env, n);
664 an = get_affinity_info(env->co, n);
667 if (n1->int_aff_neigh < 0)
668 n1->int_aff_neigh = count_interfering_aff_neighs(env, an);
670 /* build the affinity edges */
671 co_gs_foreach_neighb(an, neigh) {
672 const ir_node *m = neigh->irn;
673 int m_idx = get_irn_idx(m);
675 /* record the edge in only one direction */
680 /* skip ignore nodes */
681 if (arch_irn_is_ignore(m))
687 n2 = get_co_mst_irn(env, m);
688 if (n2->int_aff_neigh < 0) {
689 affinity_node_t *am = get_affinity_info(env->co, m);
690 n2->int_aff_neigh = count_interfering_aff_neighs(env, am);
693 * these weights are pure hackery ;-).
694 * It's not chriswue's fault but mine.
696 edge.weight = neigh->costs;
697 ARR_APP1(aff_edge_t, edges, edge);
703 /* now: sort edges and build the affinity chunks */
704 len = ARR_LEN(edges);
705 qsort(edges, len, sizeof(edges[0]), cmp_aff_edge);
706 for (i = 0; i < len; ++i) {
707 DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight));
709 (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt);
712 /* now insert all chunks into a priority queue */
713 list_for_each_entry(aff_chunk_t, curr_chunk, &env->chunklist, list) {
714 aff_chunk_assure_weight(env, curr_chunk);
716 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
717 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
718 DBG((dbg, LEVEL_1, "\n"));
720 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
723 for (pn = 0; pn < ARR_LEN(env->map.data); ++pn) {
724 co_mst_irn_t *mirn = (co_mst_irn_t*)env->map.data[pn];
727 if (mirn->chunk != NULL)
730 /* no chunk is allocated so far, do it now */
731 aff_chunk_t *curr_chunk = new_aff_chunk(env);
732 aff_chunk_add_node(curr_chunk, mirn);
734 aff_chunk_assure_weight(env, curr_chunk);
736 DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id));
737 DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk);
738 DBG((dbg, LEVEL_1, "\n"));
740 pqueue_put(env->chunks, curr_chunk, curr_chunk->weight);
746 static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chunk_t *chunk)
748 pqueue_t *grow = new_pqueue();
749 ir_node const *max_node = NULL;
753 for (i = ARR_LEN(chunk->n); i != 0;) {
754 const ir_node *irn = chunk->n[--i];
755 affinity_node_t *an = get_affinity_info(env->co, irn);
758 if (arch_irn_is_ignore(irn))
762 co_gs_foreach_neighb(an, neigh)
765 if (w > max_weight) {
773 bitset_t *visited = bitset_malloc(get_irg_last_idx(env->co->irg));
775 for (i = ARR_LEN(chunk->n); i != 0;)
776 bitset_set(visited, get_irn_idx(chunk->n[--i]));
778 pqueue_put(grow, (void *) max_node, max_weight);
779 bitset_clear(visited, get_irn_idx(max_node));
781 while (!pqueue_empty(grow)) {
782 ir_node *irn = (ir_node*)pqueue_pop_front(grow);
783 affinity_node_t *an = get_affinity_info(env->co, irn);
785 if (arch_irn_is_ignore(irn))
788 assert(i <= ARR_LEN(chunk->n));
793 /* build the affinity edges */
794 co_gs_foreach_neighb(an, neigh) {
795 co_mst_irn_t *node = get_co_mst_irn(env, neigh->irn);
797 if (bitset_is_set(visited, get_irn_idx(node->irn))) {
798 pqueue_put(grow, (void *) neigh->irn, neigh->costs);
799 bitset_clear(visited, get_irn_idx(node->irn));
805 bitset_free(visited);
810 * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node.
812 static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited,
813 aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, int col)
815 waitq *nodes = new_waitq();
817 DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%u) from %+F, color %d:", chunk->id, node->irn, col));
819 /* init queue and chunk */
820 waitq_put(nodes, node);
821 bitset_set(visited, get_irn_idx(node->irn));
822 aff_chunk_add_node(chunk, node);
823 DB((dbg, LEVEL_1, " %+F", node->irn));
825 /* as long as there are nodes in the queue */
826 while (! waitq_empty(nodes)) {
827 co_mst_irn_t *n = (co_mst_irn_t*)waitq_get(nodes);
828 affinity_node_t *an = get_affinity_info(env->co, n->irn);
830 /* check all affinity neighbors */
832 co_gs_foreach_neighb(an, neigh) {
833 const ir_node *m = neigh->irn;
834 int m_idx = get_irn_idx(m);
837 if (arch_irn_is_ignore(m))
840 n2 = get_co_mst_irn(env, m);
842 if (! bitset_is_set(visited, m_idx) &&
845 ! aff_chunk_interferes(chunk, m) &&
846 node_contains(orig_chunk->n, m))
849 following conditions are met:
850 - neighbour is not visited
851 - neighbour likes the color
852 - neighbour has not yet a fixed color
853 - the new chunk doesn't interfere with the neighbour
854 - neighbour belongs or belonged once to the original chunk
856 bitset_set(visited, m_idx);
857 aff_chunk_add_node(chunk, n2);
858 DB((dbg, LEVEL_1, " %+F", n2->irn));
859 /* enqueue for further search */
860 waitq_put(nodes, n2);
866 DB((dbg, LEVEL_1, "\n"));
872 * Fragment the given chunk into chunks having given color and not having given color.
874 static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp)
876 bitset_t *visited = bitset_malloc(get_irg_last_idx(env->co->irg));
878 aff_chunk_t *best = NULL;
880 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
883 aff_chunk_t *tmp_chunk;
884 decide_func_t *decider;
888 if (bitset_is_set(visited, get_irn_idx(irn)))
891 node = get_co_mst_irn(env, irn);
893 if (get_mst_irn_col(node) == col) {
894 decider = decider_has_color;
896 DBG((dbg, LEVEL_4, "\tcolor %d wanted\n", col));
899 decider = decider_hasnot_color;
901 DBG((dbg, LEVEL_4, "\tcolor %d forbidden\n", col));
904 /* create a new chunk starting at current node */
905 tmp_chunk = new_aff_chunk(env);
906 waitq_put(tmp, tmp_chunk);
907 expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col);
908 assert(ARR_LEN(tmp_chunk->n) > 0 && "No nodes added to chunk");
910 /* remember the local best */
911 aff_chunk_assure_weight(env, tmp_chunk);
912 if (check_for_best && (! best || best->weight < tmp_chunk->weight))
916 assert(best && "No chunk found?");
917 bitset_free(visited);
922 * Resets the temporary fixed color of all nodes within wait queue @p nodes.
923 * ATTENTION: the queue is empty after calling this function!
925 static inline void reject_coloring(struct list_head *nodes)
927 co_mst_irn_t *n, *temp;
928 DB((dbg, LEVEL_4, "\treject coloring for"));
929 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
930 DB((dbg, LEVEL_4, " %+F", n->irn));
931 assert(n->tmp_col >= 0);
933 list_del_init(&n->list);
935 DB((dbg, LEVEL_4, "\n"));
938 static inline void materialize_coloring(struct list_head *nodes)
940 co_mst_irn_t *n, *temp;
941 list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) {
942 assert(n->tmp_col >= 0);
945 list_del_init(&n->list);
949 static inline void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed)
952 assert(!node->fixed);
953 assert(node->tmp_col < 0);
954 assert(node->list.next == &node->list && node->list.prev == &node->list);
955 assert(bitset_is_set(node->adm_colors, col));
957 list_add_tail(&node->list, changed);
961 static inline int is_loose(co_mst_irn_t *node)
963 return !node->fixed && node->tmp_col < 0;
967 * Determines the costs for each color if it would be assigned to node @p node.
969 static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs)
971 int *neigh_cols = ALLOCAN(int, env->n_regs);
976 for (i = 0; i < env->n_regs; ++i) {
979 costs[i].cost = bitset_is_set(node->adm_colors, i) ? node->constr_factor : REAL(0.0);
982 for (i = 0; i < node->n_neighs; ++i) {
983 co_mst_irn_t *n = get_co_mst_irn(env, node->int_neighs[i]);
984 int col = get_mst_irn_col(n);
989 costs[col].cost = REAL(0.0);
993 coeff = REAL(1.0) / n_loose;
994 for (i = 0; i < env->n_regs; ++i)
995 costs[i].cost *= REAL(1.0) - coeff * neigh_cols[i];
999 /* need forward declaration due to recursive call */
1000 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed_ones, int depth, int *max_depth, int *trip);
1003 * Tries to change node to a color but @p explude_col.
1004 * @return 1 if succeeded, 0 otherwise.
1006 static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int exclude_col, struct list_head *changed, int depth, int *max_depth, int *trip)
1008 int col = get_mst_irn_col(node);
1011 /* neighbours has already a different color -> good, temporary fix it */
1012 if (col != exclude_col) {
1014 set_temp_color(node, col, changed);
1018 /* The node has the color it should not have _and_ has not been visited yet. */
1019 if (is_loose(node)) {
1020 col_cost_t *costs = ALLOCAN(col_cost_t, env->n_regs);
1022 /* Get the costs for giving the node a specific color. */
1023 determine_color_costs(env, node, costs);
1025 /* Since the node must not have the not_col, set the costs for that color to "infinity" */
1026 costs[exclude_col].cost = REAL(0.0);
1028 /* sort the colors according costs, cheapest first. */
1029 qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost_gt);
1031 /* Try recoloring the node using the color list. */
1032 res = recolor_nodes(env, node, costs, changed, depth + 1, max_depth, trip);
1039 * Tries to bring node @p node to cheapest color and color all interfering neighbours with other colors.
1040 * ATTENTION: Expect @p costs already sorted by increasing costs.
1041 * @return 1 if coloring could be applied, 0 otherwise.
1043 static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed, int depth, int *max_depth, int *trip)
1046 struct list_head local_changed;
1049 if (depth > *max_depth)
1052 DBG((dbg, LEVEL_4, "\tRecoloring %+F with color-costs", node->irn));
1053 DBG_COL_COST(env, LEVEL_4, costs);
1054 DB((dbg, LEVEL_4, "\n"));
1056 if (depth >= recolor_limit) {
1057 DBG((dbg, LEVEL_4, "\tHit recolor limit\n"));
1061 for (i = 0; i < env->n_regs; ++i) {
1062 int tgt_col = costs[i].col;
1066 /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */
1067 if (costs[i].cost == REAL(0.0)) {
1068 DBG((dbg, LEVEL_4, "\tAll further colors forbidden\n"));
1072 /* Set the new color of the node and mark the node as temporarily fixed. */
1073 assert(node->tmp_col < 0 && "Node must not have been temporary fixed.");
1074 INIT_LIST_HEAD(&local_changed);
1075 set_temp_color(node, tgt_col, &local_changed);
1076 DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col));
1078 /* try to color all interfering neighbours with current color forbidden */
1079 for (j = 0; j < node->n_neighs; ++j) {
1083 neigh = node->int_neighs[j];
1085 if (arch_irn_is_ignore(neigh))
1088 nn = get_co_mst_irn(env, neigh);
1089 DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_col: %d, col: %d)\n",
1090 neigh, j, nn->fixed, nn->tmp_col, nn->col));
1093 Try to change the color of the neighbor and record all nodes which
1094 get changed in the tmp list. Add this list to the "changed" list for
1095 that color. If we did not succeed to change the color of the neighbor,
1096 we bail out and try the next color.
1098 if (get_mst_irn_col(nn) == tgt_col) {
1099 /* try to color neighbour with tgt_col forbidden */
1100 neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed, depth + 1, max_depth, trip);
1108 We managed to assign the target color to all neighbors, so from the perspective
1109 of the current node, every thing was ok and we can return safely.
1112 /* append the local_changed ones to global ones */
1113 list_splice(&local_changed, changed);
1117 /* coloring of neighbours failed, so we try next color */
1118 reject_coloring(&local_changed);
1122 DBG((dbg, LEVEL_4, "\tAll colors failed\n"));
1127 * Tries to bring node @p node and all its neighbours to color @p tgt_col.
1128 * @return 1 if color @p col could be applied, 0 otherwise
1130 static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, struct list_head *changed)
1132 int col = get_mst_irn_col(node);
1134 /* if node already has the target color -> good, temporary fix it */
1135 if (col == tgt_col) {
1136 DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col));
1138 set_temp_color(node, tgt_col, changed);
1143 Node has not yet a fixed color and target color is admissible
1144 -> try to recolor node and its affinity neighbours
1146 if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) {
1147 col_cost_t *costs = env->single_cols[tgt_col];
1148 int res, max_depth, trip;
1153 DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn));
1154 res = recolor_nodes(env, node, costs, changed, 0, &max_depth, &trip);
1155 DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed"));
1156 stat_ev_int("heur4_recolor_depth_max", max_depth);
1157 stat_ev_int("heur4_recolor_trip", trip);
1163 #ifdef DEBUG_libfirm
1164 if (firm_dbg_get_mask(dbg) & LEVEL_4) {
1165 if (!is_loose(node))
1166 DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col));
1168 DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn));
1169 dbg_admissible_colors(env, node);
1170 DB((dbg, LEVEL_4, ")\n"));
1179 * Tries to color an affinity chunk (or at least a part of it).
1180 * Inserts uncolored parts of the chunk as a new chunk into the priority queue.
1182 static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c)
1184 aff_chunk_t *best_chunk = NULL;
1185 int n_nodes = ARR_LEN(c->n);
1186 int best_color = -1;
1187 int n_int_chunks = 0;
1188 waitq *tmp_chunks = new_waitq();
1189 waitq *best_starts = NULL;
1190 col_cost_t *order = ALLOCANZ(col_cost_t, env->n_regs);
1197 struct list_head changed;
1199 DB((dbg, LEVEL_2, "fragmentizing chunk #%u", c->id));
1200 DBG_AFF_CHUNK(env, LEVEL_2, c);
1201 DB((dbg, LEVEL_2, "\n"));
1203 stat_ev_ctx_push_fmt("heur4_color_chunk", "%u", c->id);
1205 ++env->chunk_visited;
1207 /* compute color preference */
1208 for (pos = 0, len = ARR_LEN(c->interfere); pos < len; ++pos) {
1209 const ir_node *n = c->interfere[pos];
1210 co_mst_irn_t *node = get_co_mst_irn(env, n);
1211 aff_chunk_t *chunk = node->chunk;
1213 if (is_loose(node) && chunk && chunk->visited < env->chunk_visited) {
1214 assert(!chunk->deleted);
1215 chunk->visited = env->chunk_visited;
1218 aff_chunk_assure_weight(env, chunk);
1219 for (i = 0; i < env->n_regs; ++i)
1220 order[i].cost += chunk->color_affinity[i].cost;
1224 for (i = 0; i < env->n_regs; ++i) {
1225 real_t dislike = n_int_chunks > 0 ? REAL(1.0) - order[i].cost / n_int_chunks : REAL(0.0);
1227 order[i].cost = (REAL(1.0) - dislike_influence) * c->color_affinity[i].cost + dislike_influence * dislike;
1230 qsort(order, env->n_regs, sizeof(order[0]), cmp_col_cost_gt);
1232 DBG_COL_COST(env, LEVEL_2, order);
1233 DB((dbg, LEVEL_2, "\n"));
1235 /* check which color is the "best" for the given chunk.
1236 * if we found a color which was ok for all nodes, we take it
1237 * and do not look further. (see did_all flag usage below.)
1238 * If we have many colors which fit all nodes it is hard to decide
1239 * which one to take anyway.
1240 * TODO Sebastian: Perhaps we should at all nodes and figure out
1241 * a suitable color using costs as done above (determine_color_costs).
1243 for (i = 0; i < env->k; ++i) {
1244 int col = order[i].col;
1246 aff_chunk_t *local_best;
1249 /* skip ignore colors */
1250 if (!bitset_is_set(env->allocatable_regs, col))
1253 DB((dbg, LEVEL_2, "\ttrying color %d\n", col));
1256 good_starts = new_waitq();
1258 /* try to bring all nodes of given chunk to the current color. */
1259 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1260 const ir_node *irn = c->n[idx];
1261 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1264 assert(! node->fixed && "Node must not have a fixed color.");
1265 DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col));
1268 The order of the colored nodes is important, so we record the successfully
1269 colored ones in the order they appeared.
1271 INIT_LIST_HEAD(&changed);
1273 good = change_node_color(env, node, col, &changed);
1274 stat_ev_tim_pop("heur4_recolor");
1276 waitq_put(good_starts, node);
1277 materialize_coloring(&changed);
1282 reject_coloring(&changed);
1284 n_succeeded += good;
1285 DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, good ? "succeeded" : "failed"));
1288 /* unfix all nodes */
1289 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1290 co_mst_irn_t *node = get_co_mst_irn(env, c->n[idx]);
1294 /* try next color when failed */
1295 if (n_succeeded == 0) {
1296 del_waitq(good_starts);
1300 /* fragment the chunk according to the coloring */
1301 local_best = fragment_chunk(env, col, c, tmp_chunks);
1303 /* search the best of the good list
1304 and make it the new best if it is better than the current */
1306 aff_chunk_assure_weight(env, local_best);
1308 DB((dbg, LEVEL_3, "\t\tlocal best chunk (id %u) for color %d: ", local_best->id, col));
1309 DBG_AFF_CHUNK(env, LEVEL_3, local_best);
1311 if (! best_chunk || best_chunk->weight < local_best->weight) {
1312 best_chunk = local_best;
1315 del_waitq(best_starts);
1316 best_starts = good_starts;
1317 DB((dbg, LEVEL_3, "\n\t\t... setting global best chunk (id %u), color %d\n", best_chunk->id, best_color));
1319 DB((dbg, LEVEL_3, "\n\t\t... omitting, global best is better\n"));
1320 del_waitq(good_starts);
1324 del_waitq(good_starts);
1327 /* if all nodes were recolored, bail out */
1328 if (n_succeeded == n_nodes)
1332 stat_ev_int("heur4_colors_tried", i);
1334 /* free all intermediate created chunks except best one */
1335 while (! waitq_empty(tmp_chunks)) {
1336 aff_chunk_t *tmp = (aff_chunk_t*)waitq_get(tmp_chunks);
1337 if (tmp != best_chunk)
1338 delete_aff_chunk(tmp);
1340 del_waitq(tmp_chunks);
1342 /* return if coloring failed */
1345 del_waitq(best_starts);
1349 DB((dbg, LEVEL_2, "\tbest chunk #%u ", best_chunk->id));
1350 DBG_AFF_CHUNK(env, LEVEL_2, best_chunk);
1351 DB((dbg, LEVEL_2, "using color %d\n", best_color));
1353 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1354 const ir_node *irn = best_chunk->n[idx];
1355 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1358 /* bring the node to the color. */
1359 DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%u\n", best_color, node->irn, best_chunk->id));
1360 INIT_LIST_HEAD(&changed);
1362 res = change_node_color(env, node, best_color, &changed);
1363 stat_ev_tim_pop("heur4_recolor");
1365 materialize_coloring(&changed);
1368 assert(list_empty(&changed));
1371 /* remove the nodes in best chunk from original chunk */
1372 len = ARR_LEN(best_chunk->n);
1373 for (idx = 0; idx < len; ++idx) {
1374 const ir_node *irn = best_chunk->n[idx];
1375 int pos = nodes_bsearch(c->n, irn);
1380 len = ARR_LEN(c->n);
1381 for (idx = nidx = 0; idx < len; ++idx) {
1382 const ir_node *irn = c->n[idx];
1388 ARR_SHRINKLEN(c->n, nidx);
1391 /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */
1392 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1393 const ir_node *n = c->n[idx];
1394 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1398 /* fragment the remaining chunk */
1399 visited = bitset_malloc(get_irg_last_idx(env->co->irg));
1400 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx)
1401 bitset_set(visited, get_irn_idx(best_chunk->n[idx]));
1403 for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) {
1404 const ir_node *irn = c->n[idx];
1405 if (! bitset_is_set(visited, get_irn_idx(irn))) {
1406 aff_chunk_t *new_chunk = new_aff_chunk(env);
1407 co_mst_irn_t *node = get_co_mst_irn(env, irn);
1409 expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0);
1410 aff_chunk_assure_weight(env, new_chunk);
1411 pqueue_put(env->chunks, new_chunk, new_chunk->weight);
1415 for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) {
1416 const ir_node *n = best_chunk->n[idx];
1417 co_mst_irn_t *nn = get_co_mst_irn(env, n);
1421 /* clear obsolete chunks and free some memory */
1422 delete_aff_chunk(best_chunk);
1423 bitset_free(visited);
1425 del_waitq(best_starts);
1427 stat_ev_ctx_pop("heur4_color_chunk");
1431 * Main driver for mst safe coalescing algorithm.
1433 static int co_solve_heuristic_mst(copy_opt_t *co)
1435 unsigned n_regs = co->cls->n_regs;
1436 bitset_t *allocatable_regs = bitset_alloca(n_regs);
1441 co_mst_env_t mst_env;
1448 ir_nodemap_init(&mst_env.map, co->irg);
1449 obstack_init(&mst_env.obst);
1451 be_put_allocatable_regs(co->cenv->irg, co->cls, allocatable_regs);
1452 k = bitset_popcount(allocatable_regs);
1454 mst_env.n_regs = n_regs;
1456 mst_env.chunks = new_pqueue();
1458 mst_env.allocatable_regs = allocatable_regs;
1459 mst_env.ifg = co->cenv->ifg;
1460 INIT_LIST_HEAD(&mst_env.chunklist);
1461 mst_env.chunk_visited = 0;
1462 mst_env.single_cols = OALLOCN(&mst_env.obst, col_cost_t*, n_regs);
1464 for (i = 0; i < n_regs; ++i) {
1465 col_cost_t *vec = OALLOCN(&mst_env.obst, col_cost_t, n_regs);
1467 mst_env.single_cols[i] = vec;
1468 for (j = 0; j < n_regs; ++j) {
1470 vec[j].cost = REAL(0.0);
1474 vec[0].cost = REAL(1.0);
1477 DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name));
1479 /* build affinity chunks */
1481 build_affinity_chunks(&mst_env);
1482 stat_ev_tim_pop("heur4_initial_chunk");
1484 /* color chunks as long as there are some */
1485 while (! pqueue_empty(mst_env.chunks)) {
1486 aff_chunk_t *chunk = (aff_chunk_t*)pqueue_pop_front(mst_env.chunks);
1488 color_aff_chunk(&mst_env, chunk);
1489 DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%u) done\n", chunk->id));
1490 delete_aff_chunk(chunk);
1493 /* apply coloring */
1494 for (pn = 0; pn < ARR_LEN(mst_env.map.data); ++pn) {
1495 co_mst_irn_t *mirn = (co_mst_irn_t*)mst_env.map.data[pn];
1496 const arch_register_t *reg;
1499 irn = get_idx_irn(co->irg, pn);
1500 if (arch_irn_is_ignore(irn))
1503 /* skip nodes where color hasn't changed */
1504 if (mirn->init_col == mirn->col)
1507 reg = arch_register_for_index(co->cls, mirn->col);
1508 arch_set_irn_register(irn, reg);
1509 DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col));
1512 /* free allocated memory */
1513 del_pqueue(mst_env.chunks);
1514 obstack_free(&mst_env.obst, NULL);
1515 ir_nodemap_destroy(&mst_env.map);
1517 stat_ev_tim_pop("heur4_total");
1522 static const lc_opt_table_entry_t options[] = {
1523 LC_OPT_ENT_INT ("limit", "limit recoloring", &recolor_limit),
1524 LC_OPT_ENT_DBL ("di", "dislike influence", &dislike_influence),
1528 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4)
1529 void be_init_copyheur4(void)
1531 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1532 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
1533 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
1534 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
1535 lc_opt_entry_t *heur4_grp = lc_opt_get_grp(co_grp, "heur4");
1537 static co_algo_info copyheur = {
1538 co_solve_heuristic_mst, 0
1541 lc_opt_add_table(heur4_grp, options);
1542 be_register_copyopt("heur4", ©heur);
1544 FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4");